Technical Deep Dive: Our Vision Model for Springshot and Spirit Airlines
How engineering for frontline workers shaped a two-stage computer vision architecture that achieved immediate adoption
The best AI model is worthless if the people meant to use it find workarounds instead.
That was one of the critical constraints driving our engineers when webAI and Springshot partnered to build a fire suppression line detection model for Spirit Airlines. The fire suppression line is a hashed marker inside an aircraft cargo hold that indicates the maximum bag-loading height permitted for the fire suppression system to function properly in an emergency. As part of the complex choreography of turning an aircraft (the 30+ tasks that happen between landing and takeoff), ground crews must photograph a properly loaded cargo hold to verify compliance.
The challenge: automating this verification in real-time, at the speed of operations.
The result: a two-stage vision architecture that processes images in under one second, has observed zero false positives in production so far, and went from zero to 100% adoption across Spirit's entire fleet in just one hour.
This is the technical story of how we built it.
The Engineering Challenge: Real-World Variability at Scale
When Spirit Airlines approached Springshot with the need for automated fire suppression line validation, the goal seemed simple: classify images taken from the cargo hold with clear visibility of the fire suppression line.
But after analyzing roughly 60,000 images from Spirit’s actual operations, the true complexity emerged. “We were seeing a lot of empty cargo bay area images, non-cargo bay images taken accidentally, and also images that were out of focus, too dark, or misaligned,” explains Fatih Altay, the Machine Learning Engineer at webAI who led the model development.
This was the core reality of the dataset: drawn directly from frontline workflows, highly unbalanced, and full of real operational noise.
To address this, we developed a custom taxonomy separating image relevance (is this even the right kind of photo?) from image quality (is the photo usable for validation?), allowing the models to capture the full variability of Spirit’s workflows and ensuring each model handled a focused part of the problem.
From that taxonomy, we identified a set of operationally meaningful image scenarios that allowed the models to reliably handle the noise, imbalance, and unpredictable conditions present in real-world aviation operations.
The Architectural Innovation: Two Models Are Better Than One
Instead of building a single model to handle every type of operational image, we proposed a two-stage architecture. Sam Avila, Machine Learning Engineer at webAI, explains the insight: "We were essentially solving two different problems. The main issue was validating whether the cargo bay was loaded properly with the fire suppression line in place. But the underlying issue was ensuring we had the right quality of images to make that validation accurately."
Our solution splits the problem:
Stage 1: Image Relevance Filter
- Loaded cargo bay (images we want)
- Empty cargo bay
- Non-cargo bay (accidental photos)
- Obstructed (cargo bay with cloth barrier during loading)
Stage 2: Image Quality Assessment
- Good image (ideal, clear fire suppression line)
- Misaligned camera
- Poor lighting
- Poor focus
The Lightweight Model Decision
"We wanted to pick a model that's available on Navigator, which can run on a laptop or an iPhone without GPU support," notes Fatih. This constraint led us to select a small, lightweight classification model—a decision that initially seemed limiting but ultimately proved beneficial.
"Those small models are not the highest-performing, state of the art models we could have chosen," Fatih acknowledges. "They're optimized for speed rather than power, which is one of the reasons why we picked a two-stage architecture." By splitting the problem across two focused models, we could achieve high accuracy and high speed despite the lightweight constraint.
Performance: The Sub-Second Imperative
Aviation operations move fast. Ground crews turning an aircraft work in a precisely choreographed 45-minute sequence, and any system that slows them down gets abandoned—regardless of how accurate it might be.
Our performance requirement: complete validation in under one second, from photo capture through two sequential models to on-screen feedback. Any longer, and workers would move on before seeing results, making the system useless in practice.
What we delivered: 700-800 milliseconds per image, under one second total including network round-trip. "If it took five to ten seconds, people would have found workarounds," notes our team. "But because it's so fast, it helped adoption."
The speed enabled workers to experience AI as ambient assistance rather than workflow disruption, which is exactly what operational AI should feel like.
Deployment Architecture: Pragmatic Cloud Solution
Why Not On-Device (Yet)
While webAI typically champions on-device AI, we made a pragmatic choice for Spirit. "Their devices are not Apple devices—they all use other providers," explains Fatih. "That's why we couldn't do it on device, and why we went with the cloud option."
The current architecture:
- Ground crew takes photo on mobile device
- Image sent to cloud S3 storage
- Cloud instance picks up image and runs both models
- Results sent back to device
- Visual feedback displayed on phone application
Even with the round-trip to the cloud solution, the system's performance remains best-in-class. The sub-second response time demonstrates how well-optimized lightweight models can deliver real-time experiences even when cloud-deployed. For comparison, many cloud-based AI systems struggle with multi-second latencies that frustrate end users. Our architecture proves that thoughtful model design and efficient infrastructure can achieve the responsiveness needed for operational contexts—whether on-device or in the cloud.
Future On-Device Potential
Despite the current cloud deployment, these lightweight models are ready for edge deployment. "Since this is a lightweight model, we expect to run this easily on any Apple device with the same speed performance," confirms Fatih. The models were originally designed to be Navigator-compatible, preparing for eventual on-device deployment when the hardware landscape allows.
Integration Excellence: Ambient AI
The most impressive metric of the project was that Spirit Airlines went from 0% to 100% adoption in one hour. Workers didn't need training or adjustment. They just saw helpful visual overlays guiding them to take better photos when the model was unable to reliably validate compliance.
This seamless integration came from working within Springshot's existing workflow rather than trying to change it. We built the AI into Springshot's existing photo capture workflow as middleware—invisible to workers but providing real-time guidance through visual overlays on their screens.
When image quality issues are detected (poor lighting, misaligned angle, obstructed view, etc.), subtle prompts guide workers to adjust before moving on. This progressive enhancement approach meant zero training, zero workflow changes, and zero resistance.
Part of the Bigger Picture
For Spirit Airlines' operations teams, this single model validates a critical safety element across 500+ daily flights with perfect reliability so far. But the real transformation runs deeper: ground crews gain confidence through instant feedback, supervisors achieve comprehensive visibility without micromanagement, and the airline builds auditable safety records that scale automatically.
As this pattern extends to other compliance checks, the 45-minute aircraft turnaround becomes smoother, faster, and measurably safer.
Zero False Positives: Safety First
After processing tens of thousands of production images, the system has maintained >99% accuracy so far, meaning it has never indicated proper fire suppression line positioning when it wasn't actually correct.
This reliability stems from the two-stage architecture itself. By filtering out irrelevant images in the first stage, we reduce the complexity for the second stage, allowing it to focus solely on quality assessment of relevant cargo bay images.
Lessons for Engineering Teams
On the other side of the project, the team has identified some key insights that we plan to apply to future projects as well:
Constraints Drive Innovation: The requirement for lightweight models forced us to think creatively, leading to the two-stage architecture that ultimately improved both speed and accuracy.
Split Complex Problems: Rather than forcing one model to handle every kind of operational image, splitting relevance from quality gave each model a focused, tractable problem—critical advantage when working with noisy, unbalanced real-world datasets.
Speed Enables Adoption: Sub-second response time was critical for real-world adoption. Even the best model fails if users find it too slow to be practical.
Start Where You Are: We deployed on cloud for non-Apple devices today while building models ready for on-device deployment tomorrow. Perfect shouldn't be the enemy of good.
Beyond Fire Suppression
It’s easy to imagine how the two-stage pattern we've proven here could extend to other aviation compliance checks including:
- Cargo door seal integrity verification before departure
- Verifying ULD locks are in the proper upright position prior to departure
- Jet bridge alignment validation for passenger safety
- FOD (Foreign Object Debris) detection on tarmac areas
- Fuel panel closure and lock verification
Each would follow the same pattern: first filter for the right type of image, then assess quality or compliance.
Technical Details for Implementation
| Training Data | 60,000 images from Spirit Airlines operatins |
|---|---|
Model Selection | ResNet Family |
Inference Time | 700-800ms per image |
Total process time | <1 second including network overhead |
Deployment | Cloud (current), iOS-ready for future |
Accuracy | >99% |
The Partnership Model
The success of this project came from combining Springshot's deep aviation operations knowledge with webAI's AI engineering capabilities. Springshot understood the workflow and had access to real operational data. webAI brought the model development and integration expertise.
Neither company could have built this alone; domain expertise and AI engineering expertise needed each other. And the partnership is far from over. We’re excited to see the future innovations that turn the real-world aviation operations of today into the AI-powered workflows of tomorrow.
To learn more about Springshot and webAI’s shared vision for the future of AI in Aviation, read How AI at the Edgie is Transforming Aviation Operations.